Landslide map spot extraction method and device based on boundary perception for remote sensing image

Abstract

The application provides a remote sensing image landslide patch extraction method and device based on boundary perception, and relates to the technical field of data processing. The method comprises the following steps: constructing a sample image set based on obtained first sample images and second sample images, performing morphological processing on a landslide contour label corresponding to the first sample images to generate a boundary label graph, configuring a patch extraction unit based on a boundary deviation metric corresponding to the sample image set and the boundary label graph, performing image division on a to-be-processed remote sensing image based on different image block specifications to obtain sub-image blocks, inputting the sub-image blocks into the configured patch extraction unit to obtain sub-probability graphs corresponding to each image block specification, performing weighted fusion on each sub-probability graph to generate a comprehensive landslide probability graph, and performing conversion processing on the comprehensive landslide probability graph to obtain landslide patch data and output the landslide patch data. The method can ensure the accuracy of landslide area recognition and effectively improve the boundary extraction precision of the landslide patch.

Description

Technical Field

[0001] This invention relates to data processing technology, and more particularly to a method and apparatus for extracting landslide patches from remote sensing images based on boundary awareness. Background Technology

[0002] With the continuous advancement of remote sensing technology and the constant improvement of image resolution, remote sensing images have become an important data source for landslide geological disaster investigation, emergency monitoring, and hidden danger investigation. Quickly and accurately extracting the extent of landslide patches from remote sensing images is the core foundation for carrying out landslide risk assessment and disaster emergency response, and is of great significance for improving the ability to prevent and control geological disasters.

[0003] Currently, most existing methods for extracting landslide patches from remote sensing images use regional range matching as the core criterion, focusing only on ensuring the overall overlap between the extracted area and the actual landslide, without setting specific constraints on the landslide boundaries. This results in output landslide patches generally having problems such as blurred boundaries, breaks, protruding burrs, and internal cavities. The boundary positioning accuracy and morphological continuity are insufficient, which cannot meet the application needs of refined disaster investigation and mapping.

[0004] Therefore, how to effectively improve the boundary extraction accuracy of landslide patches while ensuring the accuracy of landslide area identification has become an urgent problem to be solved. Summary of the Invention

[0005] This invention provides a method and apparatus for extracting landslide patches from remote sensing images based on boundary awareness, which can effectively improve the boundary extraction accuracy of landslide patches while ensuring the accuracy of landslide area identification.

[0006] A first aspect of the present invention provides a method for extracting landslide patches from remote sensing images based on boundary awareness, comprising:

[0007] A sample image set is constructed based on the first and second sample images obtained. The landslide contour labels corresponding to the first sample image are morphologically processed to generate a boundary label map.

[0008] Based on the sample image set and the boundary deviation metric corresponding to the boundary label map, the parameters of the patch extraction unit are configured to obtain the configured patch extraction unit, including:

[0009] The sample images in the sample image set are input into the patch extraction unit to obtain the predicted landslide mask output by the patch extraction unit;

[0010] A first deviation metric is determined based on the difference between the predicted landslide mask and the landslide mask corresponding to the landslide outline label;

[0011] Edge extraction is performed on the predicted landslide mask to obtain a predicted boundary map. Based on the difference between the predicted boundary map and the boundary label map, a second deviation metric is determined and used as the boundary deviation metric.

[0012] Based on the first deviation metric and the boundary deviation metric, the internal parameters of the patch extraction unit are iteratively adjusted until the preset conditions are met, thus obtaining the configured patch extraction unit.

[0013] The remote sensing image to be processed is acquired, and the image is divided into sub-image blocks based on different image block specifications. The sub-image blocks are then input into the configured patch extraction unit to obtain sub-probability maps corresponding to each image block specification.

[0014] The sub-probability maps are weighted and fused to generate a comprehensive landslide probability map. This comprehensive landslide probability map is then transformed to obtain landslide patch data, which is then output.

[0015] Assign corresponding fusion weights to each of the image block specifications, and perform a weighted average of the pixel values ​​of the same geographical location in each of the sub-probability maps according to the fusion weights to generate a comprehensive landslide probability map;

[0016] Threshold segmentation is performed on the comprehensive landslide probability map to obtain a binarized landslide mask;

[0017] The binarized landslide mask is sequentially subjected to closing operation, hole filling operation, opening operation, and connected component area filtering operation to obtain a refined landslide mask;

[0018] The contours of each landslide patch in the refined landslide mask are extracted, and combined with the georeferenced information of the remote sensing image to be processed, the contour pixel coordinates are converted into geographic coordinates to obtain landslide patch data and output it.

[0019] Optionally, in one possible implementation of the first aspect, constructing the sample image set based on the acquired first and second sample images includes:

[0020] The acquired remote sensing image slice containing landslide targets and labeled with landslide annotation information is used as the first sample image. The landslide annotation information is a landslide outline label or a landslide mask label.

[0021] The acquired remote sensing image slices that do not contain the landslide target but contain land cover types that are easily confused with the landslide target are used as the second sample images;

[0022] A sample image set is constructed based on the first sample image and the second sample image;

[0023] When constructing the sample image set, the ratio of the number of first sample images to the number of second sample images is controlled to be within a preset ratio range;

[0024] When the landslide annotation information is a landslide mask label, before performing morphological processing on the landslide contour label corresponding to the first sample image, the method further includes:

[0025] Connectivity analysis was performed on the landslide mask labels to extract the contour vertex coordinates of each landslide patch;

[0026] The coordinates of the contour vertices are normalized to obtain the landslide contour label.

[0027] Optionally, in one possible implementation of the first aspect, the step of performing morphological processing on the landslide contour label corresponding to the first sample image to generate a boundary label map includes:

[0028] The polygonal region enclosed by the coordinates of each vertex in the landslide outline label is filled with pixels to generate a landslide mask;

[0029] A gradient convolution kernel is applied to the landslide mask to calculate the gradient magnitude at each pixel location, and a boundary label map is generated after normalization.

[0030] Alternatively, an expansion operation can be performed on the landslide mask to obtain an expanded mask, and an corrosion operation can be performed on the landslide mask to obtain a corrosion mask;

[0031] The boundary label map is obtained by subtracting the expansion mask from the corrosion mask.

[0032] Optionally, in one possible implementation of the first aspect, determining the first deviation metric based on the difference between the predicted landslide mask and the landslide mask corresponding to the landslide profile label includes:

[0033] Obtain the predicted target box output by the patch extraction unit, and determine the target box deviation sub-metric based on the positional difference between the predicted target box and the real target box corresponding to the landslide contour label;

[0034] Obtain the predicted category confidence score output by the patch extraction unit, and determine the category bias sub-measure based on the difference between the predicted category confidence score and the true category label;

[0035] Obtain the regional overlap difference between the predicted landslide mask and the landslide mask corresponding to the landslide outline label, and determine the mask deviation sub-metric;

[0036] The target box deviation sub-metric, the category deviation sub-metric, and the mask deviation sub-metric are weighted and calculated to obtain the first deviation metric.

[0037] Optionally, in one possible implementation of the first aspect, determining the second deviation metric based on the difference between the predicted boundary map and the boundary label map includes:

[0038] The predicted values ​​of each pixel in the predicted boundary map are compared with the actual values ​​of the corresponding pixels in the boundary label map to obtain a pixel difference measure between the predicted boundary map and the boundary label map.

[0039] Calculate the overlap ratio between the predicted boundary map and the boundary label map to obtain the boundary overlap metric;

[0040] A second deviation metric is determined based on the pixel difference metric and the boundary coincidence metric.

[0041] Optionally, in one possible implementation of the first aspect, the first deviation measure and the boundary deviation measure are weighted and combined to obtain a comprehensive deviation measure. Based on the comprehensive deviation measure, the internal parameters of the patch extraction unit are adjusted to obtain an adjusted patch extraction unit.

[0042] The sample images in the sample image set are input again into the adjusted patch extraction unit to calculate a new comprehensive deviation metric.

[0043] When the change between the new comprehensive deviation metric and the comprehensive deviation metric before one round of adjustment is less than a preset threshold, the iterative adjustment is stopped, and the current patch extraction unit is used as the configured patch extraction unit.

[0044] If the change is greater than or equal to a preset threshold, the next round of parameter adjustment will continue based on the new comprehensive deviation metric until the change meets the preset threshold.

[0045] Optionally, in one possible implementation of the first aspect, the step of acquiring the remote sensing image to be processed, dividing the remote sensing image to be processed into sub-image blocks based on different image block specifications, and inputting the sub-image blocks into the configured patch extraction unit to obtain sub-probability maps corresponding to each image block specification includes:

[0046] At least two different image block specifications are preset. For each image block specification, the remote sensing image to be processed is divided based on the corresponding image block division method. There are overlapping areas between adjacent sub-image blocks, and a set of sub-image blocks corresponding to the image block specification is obtained.

[0047] The sub-image block sets corresponding to each of the image block specifications are respectively input into the configured image patch extraction unit to obtain the landslide prediction results of each of the sub-image blocks;

[0048] The landslide prediction results of each sub-image block belonging to the same image block size are stitched together according to the original position of each sub-image block in the remote sensing image to be processed to obtain the sub-probability map corresponding to each image block size.

[0049] A second aspect of the present invention provides a boundary-aware remote sensing image landslide patch extraction device according to any one of claims 1-7, comprising:

[0050] The construction module is used to construct a sample image set based on the acquired first sample image and second sample image, and to perform morphological processing on the landslide contour label corresponding to the first sample image to generate a boundary label map.

[0051] The configuration module is used to configure the parameters of the patch extraction unit based on the sample image set and the boundary deviation metric corresponding to the boundary label map, so as to obtain the configured patch extraction unit.

[0052] The segmentation module is used to acquire the remote sensing image to be processed, divide the remote sensing image to be processed into sub-image blocks based on different image block specifications, and input the sub-image blocks into the configured patch extraction unit to obtain sub-probability maps corresponding to each image block specification.

[0053] The conversion module is used to perform weighted fusion of the sub-probability maps to generate a comprehensive landslide probability map, and to perform conversion processing on the comprehensive landslide probability map to obtain landslide patch data and output it.

[0054] The beneficial effects of this invention are as follows:

[0055] 1. This invention constructs a sample image set containing a first sample image and a second sample image. The first sample image is a positive sample containing landslides, and the second sample image is a negative sample containing easily confused features. During the construction process, the ratio of the two types of samples is controlled within a preset range. This allows the patch extraction unit to learn the sample features of landslides and also access the non-landslide features of easily confused features such as bare land, quarries, and road slopes during parameter configuration. This reduces the false detection rate while avoiding the increase in the false detection rate caused by too many negative samples, thereby improving the recognition accuracy.

[0056] 2. This invention generates a boundary label map by performing morphological processing on landslide contour labels, and introduces a boundary deviation metric based on this boundary label map during the parameter configuration process of the patch extraction unit. This enables the configured patch extraction unit to more accurately delineate the landslide boundary contour while identifying the landslide area, thereby improving the boundary integrity and positioning accuracy of the landslide patches.

[0057] 3. This invention divides the remote sensing image to be processed using at least two different image block sizes, which can take into account the extraction needs of landslide targets at different scales and ensure the integrity of the extraction results. This invention can also transform the fused comprehensive landslide probability map to make the output landslide patch boundaries continuous, complete, and clearly defined. Attached Figure Description

[0058] Figure 1 This is a schematic flowchart of a method for extracting landslide patches from remote sensing images based on boundary awareness, provided in an embodiment of the present invention.

[0059] Figure 2 This is a schematic diagram of the internal structure of a patch extraction unit provided in an embodiment of the present invention;

[0060] Figure 3 This is a schematic diagram of the construction process of boundary-aware constraints provided in an embodiment of the present invention;

[0061] Figure 4 This is a schematic diagram of a landslide patch extraction device based on boundary awareness from remote sensing images provided in an embodiment of the present invention. Detailed Implementation

[0062] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0063] See Figure 1 This is a schematic flowchart of a method for extracting landslide patches from remote sensing images based on boundary awareness, provided in an embodiment of the present invention. Figure 1 The method shown can be implemented by software and / or hardware devices. It includes steps S1 to S4, as detailed below:

[0064] S1. Construct a sample image set based on the acquired first sample image and second sample image, perform morphological processing on the landslide contour label corresponding to the first sample image, and generate a boundary label map.

[0065] The first sample image refers to a remote sensing image slice containing landslide targets and labeled with landslide annotation information; it can also be called a positive sample. The second sample image refers to a remote sensing image slice that does not contain landslide targets but contains land cover types that are easily confused with landslides; it can also be called a negative sample. A landslide target refers to an actual surface damage area in the remote sensing image formed by the downward sliding of slope rock and soil along a weak surface or weak zone, either wholly or partially. Landslide annotation information refers to data formed by manually marking the location and boundaries of landslide targets in the first sample image; it is used to indicate the actual landslide area that the patch extraction unit should refer to during parameter configuration. This image set refers to the image collection composed of the first sample image and the second sample image. Remote sensing image slices refer to local image regions cropped from a large remote sensing image at a certain size, whose size matches the image size that the patch extraction unit can directly process. Landslide contour labels refer to the annotation data recorded in the form of polygon vertex coordinate sequences, used to describe the boundary position and shape of landslide patches. Morphological processing refers to the processing method of extracting specific structural regions by performing shape-level operations on landslide contour labels. Boundary label map refers to the reference image that retains only the landslide boundary position information and is used to indicate the location of the landslide boundary.

[0066] Understandably, in remote sensing imagery, landforms such as bare land, quarries, roadside slopes, and riverbanks are highly similar to landslides in spectral and textural features. If only the first sample image is used to configure the parameters of subsequent patch extraction units, these units will only encounter landslide sample features during the configuration process, without encountering sample features of easily confused landforms. In practical applications, these easily confused landforms are easily misclassified as landslides, leading to an increased false detection rate. Similarly, simply adding a large number of second sample images may cause the unit to encounter too many non-landslide samples during the configuration process, making the patch extraction unit's judgment of landslides more conservative and missing real landslides with less obvious features, leading to an increased false detection rate. Therefore, the first and second sample images can be combined to construct a sample image set, and the ratio of their quantities should be controlled during the construction process so that the subsequent parameter configuration can achieve a balance between reducing false detections and avoiding false detections.

[0067] Specifically, when constructing the sample image set, the remote sensing image slices that contain landslide targets and are labeled with landslide annotation information can be used as the first sample image, and the remote sensing image slices that do not contain landslide targets but contain land cover types that are easily confused with landslides can be used as the second sample image. The two types of sample images are combined to complete the construction of the sample image set. When generating the boundary label map, the landslide contour labels corresponding to the first sample image are morphologically processed, and the area that only represents the landslide boundary location is extracted from the landslide contour labels. The extraction result is used as the boundary label map.

[0068] In some embodiments, step S1, "constructing a sample image set based on the acquired first and second sample images," includes the following steps:

[0069] S11, the acquired remote sensing image slice containing the landslide target and labeled with landslide annotation information is used as the first sample image, wherein the landslide annotation information is a landslide outline label or a landslide mask label.

[0070] Among them, landslide mask labels refer to the annotation data recorded in the form of binary images to identify the pixel area where the landslide target is located. Pixels in the image that belong to the landslide area are assigned a first value, and pixels that do not belong to the landslide area are assigned a second value. The first value and the second value are different.

[0071] Understandably, in actual landslide annotation work, different annotators or different annotation tools may use different annotation formats. Some annotation data directly records the boundary contour of the landslide in the form of polygon vertex coordinates, while others record the pixel area occupied by the landslide in the form of a mask image. Therefore, landslide annotation information can include both landslide contour labels and landslide mask labels, making the source of sample images more flexible and compatible with data generated by different annotation methods.

[0072] Specifically, when collecting the first sample images, for each remote sensing image slice containing the landslide target, the format of its attached annotation information is checked. If the annotation information is a sequence of polygon vertex coordinates, it is directly used as the landslide contour label. If the annotation information is a binary image, it is used as the landslide mask label. Both formats can be used as valid landslide annotation information for the construction of the sample image set.

[0073] S12, the acquired remote sensing image slices that do not contain the landslide target but contain land cover types that are easily confused with the landslide target are used as the second sample images.

[0074] Among them, land cover type refers to the classification name of different land cover or land use types in remote sensing images. Land cover types that are easily confused with landslide targets refer to land cover types that are similar to landslides in spectral and textural features, including bare land, road slopes, quarries, river beaches, construction sites, etc.

[0075] Understandably, landslides in remote sensing images typically appear as grayish-brown or light-colored areas with rough surfaces and sparse vegetation. These features are very similar to features such as bare land, quarries, and construction sites. If the sample image set only contains the first sample image, the patch extraction unit can only learn the features of landslides during parameter configuration and has never encountered these easily confused features. After configuration, in actual use, once it encounters areas such as quarries or construction sites, the unit is likely to misclassify them as landslides. Therefore, remote sensing image slices containing these easily confused features but not landslides can be collected as second sample images, allowing the patch extraction unit to encounter the features of these easily confused features during parameter configuration, thereby learning to distinguish landslides from similar features.

[0076] Specifically, when collecting the second sample images, remote sensing images of the study area or adjacent areas are selected to include easily confused landform types such as bare land, road slopes, quarries, river beaches, construction sites, exposed farmland, bare rock areas, shaded areas, or water edges, which do not contain landslide targets. After confirming that there are indeed no landslide targets in each slice, it is used as the second sample image, and an empty annotation file with the same name is created for it to indicate that the image does not contain landslides.

[0077] S13, construct a sample image set based on the first sample image and the second sample image.

[0078] Specifically, the determined first sample image and the determined second sample image are combined and organized into a unified sample image set. Each image in the sample image set corresponds to a labeling file. The labeling file corresponding to the first sample image contains landslide labeling information, while the labeling file corresponding to the second sample image is empty.

[0079] S14, when constructing the sample image set, the ratio of the number of the first sample image to the number of the second sample image is controlled to be within a preset ratio range.

[0080] The quantity ratio refers to the ratio between the total number of the first sample images and the total number of the second sample images in the sample image set, and the preset ratio range refers to the pre-defined range of the quantity ratio between the first sample images and the second sample images.

[0081] It is understandable that by controlling the ratio of the number of first sample images and second sample images within a reasonable preset range, the parameter configuration process can achieve a balance between reducing false detections and avoiding false negatives. The preset ratio range can be adjusted according to the tolerance for false detections and false negatives in the actual application scenario. When the tolerance for false detections is low, the ratio of the second sample images can be appropriately increased, and when the tolerance for false negatives is low, the ratio of the second sample images can be appropriately decreased.

[0082] Specifically, when the first sample image and the second sample image are combined into a sample image set, the number of each type of sample image is counted, and the ratio between them is calculated. If the ratio is not within the preset ratio range, the number of second sample images included can be adjusted to make the final ratio fall within the preset ratio range. The adjustment method can be to randomly select some of the second sample images that have been collected, or to collect more samples of a certain type from the candidate images.

[0083] In some embodiments, when the landslide annotation information is a landslide mask label, before performing morphological processing on the landslide contour label corresponding to the first sample image, the method further includes:

[0084] A1. Perform connected component analysis on the landslide mask labels to extract the contour vertex coordinates of each landslide patch.

[0085] Connected component analysis refers to the process of merging pixels with the same pixel value and adjacent spatial positions in a binary image into the same region. Each merged independent region corresponds to a landslide patch. A landslide patch is a pixel region occupied by a complete and independent landslide target in a remote sensing image, consisting of a group of adjacent pixels. Contour vertex coordinates refer to a set of pixel coordinate points extracted sequentially along the outer boundary of the landslide patch, used to describe the outer contour shape of the patch.

[0086] Understandably, a landslide mask label is a binary image where all pixels belonging to a landslide are labeled with the same value. However, it cannot directly distinguish which of these pixels belong to which different landslide individuals. If there are multiple unconnected landslides in a remote sensing image slice, they will be uniformly labeled as landslide pixels in the mask label, but they are not connected to each other. Through connected component analysis, pixels with the same value and adjacent positions can be grouped into an independent region. Each independent connected region is a separate landslide patch. After distinguishing different landslide patches, the coordinates of the outer contour vertices of each patch can be extracted to generate an independent polygon label for each landslide instance.

[0087] Specifically, when the landslide annotation information of the first sample image is a landslide mask label, connected component analysis is performed on the mask label. Each pixel in the mask image is scanned, and pixels whose pixel values ​​represent landslides and are spatially adjacent are merged into the same connected region. Each connected region corresponds to an independent landslide patch. For each landslide patch, the pixel coordinates along its outer boundary are tracked and recorded, and arranged in order to form the contour vertex coordinate sequence of the patch. The number of sets of contour vertex coordinates is the same as the number of landslide patches.

[0088] A2, normalize the coordinates of the contour vertices to obtain the landslide contour label.

[0089] Normalization refers to the process of dividing the x-coordinate of each vertex in the contour vertex coordinates by the width of the remote sensing image slice and the y-coordinate by the height of the remote sensing image slice, so that the coordinate values ​​are mapped from pixel units to a range of 0 to 1.

[0090] Understandably, the extracted contour vertex coordinates are absolute coordinate values ​​in pixels, and their magnitude depends on the specific size of the remote sensing image slice. For example, the same landslide contour might have vertex coordinates of (320, 256) in a 640×640 slice, while the vertex coordinates at the same location in a 1024×1024 slice would become (512, 410). Through normalization, the pixel coordinates are converted into proportional values ​​relative to the image width and height. Regardless of the image slice size, the normalized coordinates are between 0 and 1, allowing the annotation data in slices of different sizes to be processed uniformly.

[0091] Specifically, for each set of extracted contour vertex coordinates, each vertex on the contour is traversed, and the x-coordinate value of the vertex is divided by the width of the remote sensing image slice, and the y-coordinate value is divided by the height of the remote sensing image slice to obtain normalized coordinate values. The normalized contour vertex coordinates of each set are organized according to a preset label format to obtain landslide contour labels corresponding to the first sample image. Each landslide instance in the label corresponds to a row of data, which includes a category number and a set of normalized polygon vertex coordinates.

[0092] In some embodiments, step S1, "performing morphological processing on the landslide contour label corresponding to the first sample image to generate a boundary label map," includes the following steps:

[0093] S15, fill the polygonal area enclosed by the coordinates of each vertex in the landslide outline label with pixels to generate a landslide mask.

[0094] The landslide mask refers to a binary image with the same size as a slice of remote sensing image. Pixels inside the polygon bounded by the vertex coordinates of the landslide outline label can be assigned a first value representing a landslide, while pixels outside the polygon can be assigned a second value representing a non-landslide. The first and second values ​​are different.

[0095] Understandably, landslide contour labels are vector data recorded in the form of polygon vertex coordinate sequences, describing the geometry of the landslide boundary. However, they are not images themselves and cannot be directly subjected to image-level morphological operations. Whether using gradient convolution kernels to extract edges or using dilation and erosion operations to extract boundary loops, these operations need to be performed on a raster image. Therefore, the vector polygon contour can be converted into a raster binary image, and the inner region of the polygon can be filled with one value and the outer region can be filled with another value to generate a landslide mask.

[0096] Specifically, create a blank image with the same size as the remote sensing image slice. Set the initial value of all pixels in the image to the second value representing non-landslide. Traverse the polygonal area enclosed by the vertex coordinates of each vertex in the landslide outline label, and modify the value of all pixels inside the polygon to the first value representing landslide. After completing the pixel filling, the landslide mask is obtained.

[0097] S16, apply a gradient convolution kernel to the landslide mask, calculate the gradient magnitude at each pixel position, and generate a boundary label map after normalization.

[0098] The gradient convolution kernel is a calculation template used to detect the degree of drastic change in pixel values ​​in an image. By sliding the template pixel by pixel on the image and performing weighted summation, areas with significant gray-scale changes in the image can be highlighted. The gradient magnitude refers to the numerical value of the intensity of pixel value change calculated by the gradient convolution kernel at each pixel position in the image. The more drastic the change, the larger the gradient magnitude.

[0099] Understandably, in a landslide mask, pixel values ​​are the same within the landslide area and also within the non-landslide area. Only at the boundary between the landslide and non-landslide areas do pixel values ​​change abruptly. The gradient convolution kernel detects the location of such abrupt changes in pixel values. The more drastic the change, the larger the calculated gradient magnitude. Therefore, at the landslide boundary, the gradient magnitude will be significantly higher than in other areas. By normalizing the gradient magnitude, it can be mapped to a uniform numerical range, making the numerical differences between the boundary and non-boundary locations more consistent. This generates a boundary label map that only identifies the boundary location. This method performs gradient calculations directly on the landslide mask without additional dilation and erosion operations, resulting in finer boundary lines and more precise locations in the generated boundary label map.

[0100] Specifically, for the generated landslide mask, pixel-by-pixel convolution operations are performed using both horizontal and vertical gradient convolution kernels to obtain the pixel value changes at each pixel location in the horizontal and vertical directions. For each pixel location, the gradient magnitude at that location is calculated based on the horizontal and vertical changes, using the following formula:

[0101]

[0102] in, The horizontal gradient convolution kernel, The vertical gradient convolution kernel, This represents the convolution operation. For the input landslide mask, To avoid extremely small positive numbers with unstable values, This is the normalization function.

[0103] The calculated gradient magnitude map is normalized to map the magnitude to a preset numerical range, generating a boundary label map. In the boundary label map, pixels at landslide boundary locations have higher values, while pixels at non-boundary locations have lower values.

[0104] S17, perform an expansion operation on the landslide mask to obtain an expanded mask, and perform an corrosion operation on the landslide mask to obtain a corrosion mask.

[0105] Among them, dilation operation refers to the process of expanding the pixels belonging to the landslide area in a binary image outward by one ring, so that the range of the landslide area is slightly enlarged. Dilation mask refers to the binary image obtained after dilation operation with the landslide area slightly enlarged. Erosion operation refers to the process of shrinking the pixels belonging to the landslide area in a binary image inward by one ring, so that the range of the landslide area is slightly reduced. Erosion mask refers to the binary image obtained after erosion operation with the landslide area slightly reduced.

[0106] It is understandable that dilation and erosion are two morphological operations in opposite directions. Dilation expands the landslide area outward, making the landslide area larger than the original area, while erosion shrinks the landslide area inward, making the landslide area smaller than the original area. Both deviate from the original landslide boundary from the outside and inside, respectively, preparing for subsequent subtraction to extract the boundary zone. Extracting the boundary using gradient convolution kernels and extracting the boundary using dilation and erosion subtraction are two optional methods, and one can be selected to generate the boundary label map according to actual needs.

[0107] Specifically, in the dilation operation, both the dilation and erosion operations rely on a small template called a structuring element. This template is typically a small rectangular or circular area, such as a 3×3 pixel square. During the dilation operation, the structuring element is moved pixel by pixel on the slippery mask. At each position, it is checked whether there is at least one slippery pixel in the area covered by the structuring element. If there is, the pixel at the center of the structuring element is set as a slippery pixel. If no slippery pixels exist, they are set as non-slippery pixels. After traversing the entire slippery mask, those non-slippery pixels that were originally adjacent to the boundary of the slippery region are also marked as slippery pixels because the structuring element covers the adjacent slippery pixels. The slippery region expands outward by a ring, thus obtaining the dilation mask.

[0108] During the erosion operation, the structuring element is moved pixel by pixel on the landslide mask. However, unlike the dilation operation, the erosion operation requires that all pixels within the area covered by the structuring element must be landslide pixels in order to retain the pixel at the center of the structuring element as a landslide pixel. If any non-landslide pixel exists within the covered area, the pixel at the center is set as a non-landslide pixel. After traversing the entire landslide mask, the pixels located at the edge of the landslide area are removed because the structuring element covers the non-landslide pixels outside the area, which does not meet the condition of being all landslide pixels. The landslide area shrinks inward by one ring, thus obtaining the erosion mask.

[0109] For example, suppose we have a landslide mask with a roughly circular landslide area in the center. After performing a dilation operation using a 3×3 square structuring element, this landslide area will be larger than before. This is because the non-landslide pixels that were originally adjacent to the landslide boundary are now detected as landslide pixels under the structuring element's coverage and are therefore marked as landslide pixels. However, after performing an erosion operation using the same 3×3 structuring element, this landslide area will be smaller than before. This is because the pixels located at the edge of the original landslide area will always contain non-landslide pixels within the area covered by the surrounding structuring element, failing to meet the condition of being entirely landslide pixels, and are therefore removed. Only the pixels in the central area are retained. This example shows that dilation makes the landslide area larger, while erosion makes it smaller. The results of the two methods show a significant difference at the boundary, with a ring-shaped area present in the dilated mask but not in the eroded mask.

[0110] S18, perform a subtraction operation between the expansion mask and the corrosion mask to obtain a boundary label map.

[0111] Understandably, the landslide area in the expansion mask has expanded outwards, making it slightly larger than the original landslide area, while the landslide area in the corrosion mask has contracted inwards, making it slightly smaller than the original landslide area. Subtracting the two pixel by pixel, the area present in the expansion mask but absent in the corrosion mask is precisely a closed loop near the original landslide boundary. This loop surrounds the original landslide boundary, accurately identifying its location. Unlike the fine-line boundary extracted by the gradient convolution kernel, the boundary label map obtained by the difference between expansion and corrosion is a boundary band with a certain width, the width of which is determined by the size of the structuring element used in the expansion and corrosion operations.

[0112] Specifically, the values ​​of corresponding pixel positions in the obtained dilatation mask and erosion mask are subtracted pixel by pixel. The landslide area in the dilatation mask is larger than the landslide area in the erosion mask. After subtraction, the pixel areas present in the dilatation mask but not in the erosion mask form a closed ring-shaped area around the original landslide boundary. This ring-shaped area is the boundary label map. Pixels within the ring-shaped area can be assigned a first value, and the remaining pixels can be assigned a second value.

[0113] S2, based on the boundary deviation metric corresponding to the sample image set and the boundary label map, the parameters of the patch extraction unit are configured to obtain the configured patch extraction unit.

[0114] Among them, the boundary deviation metric refers to a numerical index that quantifies the degree of difference between the predicted boundary output by the patch extraction unit and the boundary label map. The smaller the value, the closer the predicted boundary is to the real boundary. The patch extraction unit is a processing unit used to identify and extract landslide patches from remote sensing images. It contains a set of adjustable parameters. Different parameter values ​​determine the unit's ability to identify landslides and process boundaries. The patch extraction unit can be implemented using an instance segmentation network that can output target boxes, category confidence scores, and target masks, such as the YOLO11-seg segmentation network. After parameter configuration, the unit has the ability to identify landslide areas from remote sensing images and output landslide masks. Parameter configuration refers to the process of repeatedly adjusting the internal parameters of the patch extraction unit to make its output results gradually approach the real labels.

[0115] See Figure 2 This is a schematic diagram of the internal structure of a feature extraction unit provided in an embodiment of the present invention. The internal structure and constraint mechanism of the feature extraction unit are as follows: Figure 2As shown, the overall architecture adopts a hierarchical structure, consisting of the following modules from input to output: a main feature extraction module, used to extract multi-level features from the input remote sensing image and output three sets of feature maps with different resolutions, capturing multi-dimensional image features from fine-grained texture details to large-scale semantic information layer by layer; a multi-scale feature fusion module, which performs cross-scale information interaction on feature maps at different levels through upsampling, feature stitching, and downsampling operations, and outputs three sets of optimized feature maps adapted to small, medium, and large targets to cope with the significant differences in landslide scale; a detection and segmentation output unit, which sets detection and segmentation branches for the three sets of scale features respectively, with the detection branch outputting the predicted bounding box and class confidence of the landslide target, and the segmentation branch outputting the pixel-level prediction mask of the landslide area; and a boundary constraint auxiliary branch, used to extract predicted boundary features from the prediction mask and compare them with the pre-generated real boundary labels, outputting a boundary deviation metric.

[0116] Understandably, before parameter configuration, the internal parameters of the patch extraction unit are in an initial state. At this point, the unit does not have the ability to accurately identify landslide areas and precisely delineate landslide boundaries from remote sensing images. To enable the unit to acquire this ability, it is necessary to repeatedly adjust the unit's internal parameters using a constructed sample image set and boundary label map. During the adjustment process, on the one hand, it is necessary to measure the difference between the landslide area predicted by the unit and the actual landslide area, so that the unit gradually learns to accurately identify the location and extent of the landslide. On the other hand, it is also necessary to measure the difference between the landslide boundary predicted by the unit and the boundary label map, that is, to introduce a boundary deviation metric, so that the unit gradually learns to accurately delineate the landslide boundary. Boundary contour: If only regional difference measurement is used during parameter configuration, the unit will only optimize in the direction of reducing the overall regional difference when adjusting parameters, and the accuracy of the boundary position will not be improved. The final output patch boundary may still have problems with discontinuity or inaccuracy. After using the boundary deviation measurement as an independent basis for parameter adjustment, the unit must take into account both regional accuracy and boundary accuracy during parameter configuration. Parameter configuration is an iterative process. After each round of adjustment, it is necessary to re-examine whether the deviation measurement has decreased, so as to determine whether the parameters are being optimized in the right direction. When the change of the deviation measurement approaches stability, it means that the parameters of the unit have been basically adjusted.

[0117] Specifically, the sample images in the sample image set are input into the patch extraction unit. The output result of the patch extraction unit is checked using the boundary deviation metric corresponding to the constructed boundary label map. The internal parameters of the unit are repeatedly adjusted according to the check result. After each round of adjustment, the adjustment effect is checked again using sample images until the preset conditions are met. The patch extraction unit at this time is then used as the configured patch extraction unit.

[0118] Specifically, sample images from the sample image set are input into the patch extraction unit to obtain the predicted landslide mask output by the unit. Based on the difference between the predicted landslide mask and the actual landslide mask corresponding to the landslide contour label, a first deviation metric is determined. Simultaneously, based on the difference between the predicted landslide mask and the boundary label map, a boundary deviation metric is determined. Based on the first deviation metric and the boundary deviation metric, the internal parameters of the patch extraction unit are adjusted. After adjustment, the sample images are input into the unit again to recalculate the deviation metric. The above adjustment and verification process is repeated until the change in the deviation metric meets the preset conditions. The patch extraction unit at this point is then used as the configured patch extraction unit.

[0119] Based on the above embodiments, step S2 can be implemented in the following ways:

[0120] S21, input the sample images from the sample image set into the patch extraction unit to obtain the predicted landslide mask output by the patch extraction unit.

[0121] Among them, the sample image refers to the remote sensing image slice used for parameter adjustment in this round of sample image set, which can be the first sample image or the second sample image. The predicted landslide mask refers to the binary image or probability image output by the patch extraction unit after processing the input image according to the current internal parameters, which represents the distribution of landslide areas predicted by the unit. The value of each pixel in the image represents the probability or category that the unit judges that the location belongs to a landslide.

[0122] Specifically, a batch of sample images are taken from the sample image set and input into the patch extraction unit one by one. For each sample image, the patch extraction unit performs pixel-by-pixel analysis on the image based on its current internal parameters to determine whether each pixel position belongs to the landslide area. Finally, a predicted landslide mask with the same size as the input image is output. The value of each pixel in the mask reflects the unit's judgment result that the position belongs to the landslide.

[0123] S22, Based on the difference between the predicted landslide mask and the landslide mask corresponding to the landslide contour label, a first deviation metric is determined.

[0124] The first deviation metric is a numerical indicator that quantifies the difference in regional coverage between the predicted landslide mask output by the patch extraction unit and the landslide mask generated by the landslide outline label.

[0125] It is understandable that there is a certain difference between the prediction results of the patch extraction unit and the actual annotation. The first deviation metric can measure the degree of consistency between the predicted landslide area and the actual landslide area in terms of overall coverage. The more the predicted area overlaps with the actual area and the less the deviation, the smaller the first deviation metric, indicating that the unit is more accurate in judging the location and extent of the landslide area.

[0126] Specifically, the predicted landslide mask can be compared with the landslide mask generated from the landslide contour label. The two can be compared pixel by pixel to see if their judgments of the same location are consistent. The overall difference between the prediction and the actual label can be statistically analyzed. Based on the comparison results, a comprehensive numerical index can be calculated as the first deviation measure.

[0127] Based on the above embodiments, step S22 can be implemented in the following ways:

[0128] S221, Obtain the predicted target box output by the patch extraction unit, and determine the target box deviation sub-metric based on the positional difference between the predicted target box and the real target box corresponding to the landslide contour label.

[0129] Among them, the predicted target box refers to the rectangular bounding box output by the patch extraction unit when it outputs the predicted landslide mask and locates the area where the landslide target is located. It is used to identify the approximate location range of the landslide target in the image as determined by the unit. The true target box refers to the smallest bounding rectangle that can exactly enclose the landslide contour, calculated based on the coordinates of each vertex in the landslide contour label. It is used to identify the location range of the true landslide target in the image. The position difference refers to the degree of deviation between the predicted target box and the true target box in geometric parameters such as center coordinates, width, and height. The target box deviation sub-metric is a numerical index obtained by quantifying the position difference between the predicted target box and the true target box. The smaller the value, the more accurate the localization.

[0130] Specifically, the x and y coordinates of the center point of the predicted target box in the image, as well as the width and height of the target box, can be obtained from the output of the patch extraction unit. The corresponding parameters can be obtained from the real target box corresponding to the landslide contour label. The differences between the predicted target box and the real target box in the x-coordinate of the center point, the y-coordinate of the center point, the width, and the height are calculated respectively. These four differences are then combined according to a preset mathematical formula, such as taking the sum of squares or weighted sums of the differences. The result is the target box deviation metric.

[0131] S222, obtain the predicted category confidence score output by the patch extraction unit, and determine the category bias sub-metric based on the difference between the predicted category confidence score and the true category label.

[0132] Among them, the predicted category confidence score refers to the probability value given by the patch extraction unit for the possibility that the target in the predicted target box belongs to the landslide category. The larger the value, the more confident that the target in the box is a landslide. The true category label refers to the true category of the target marked in the sample image. For the landslide target in the first sample image, its true category label is landslide. For the second sample image, its true category label is not landslide. The category bias sub-metric is a value that reflects the degree of inconsistency between the predicted category confidence score and the true category label. The larger the value, the greater the deviation between the prediction and the reality.

[0133] Understandably, the patch extraction unit not only needs to know where the landslide is, but also needs to accurately determine whether the detected target is actually a landslide. If the confidence level of the unit's category judgment deviates significantly from the actual category, such as classifying a bare area as a landslide with a high confidence level, it indicates that the unit's category judgment ability is insufficient. By measuring the difference between the predicted category confidence level and the true category label, the parameter adjustment process can be guided to optimize the unit's category judgment ability, making it more accurate in distinguishing between landslides and non-landslides.

[0134] Specifically, the predicted class confidence score for each predicted target is obtained from the output of the patch extraction unit, along with the true class label for that target. For targets whose true class is landslide, the predicted class confidence score of the unit should be as close to 1 as possible. In this case, the class bias of the target is the difference between the predicted class confidence score and 1. For targets whose true class is not landslide, the predicted class confidence score of the unit should be as close to 0 as possible. In this case, the class bias of the target is the predicted class confidence score itself. The class biases of all predicted targets are summarized according to a preset mathematical method, and the result obtained after summarizing is the class bias sub-metric.

[0135] S223, obtain the regional overlap difference between the predicted landslide mask and the landslide mask corresponding to the landslide contour label, and determine the mask deviation sub-metric.

[0136] Among them, the regional overlap difference refers to the degree of inconsistency in spatial coverage between the pixel regions marked as landslides in the predicted landslide mask and the pixel regions marked as landslides in the landslide mask generated by the landslide contour label. This includes redundant pixels that are predicted as landslides but are not actually landslides, and missing pixels that are actually landslides but are not predicted as landslides. The mask deviation sub-metric is the value obtained after comprehensively calculating the regional overlap difference.

[0137] Understandably, the bounding box provides the approximate location range of the landslide, and the category confidence provides the probability of belonging to a landslide, but neither can reflect the specific shape and boundary of the landslide area. The mask deviation sub-metric can measure the degree of overlap between the predicted landslide mask and the real landslide mask at the pixel level. This sub-metric can reflect the ability of the patch extraction unit to finely delineate the shape and range of the landslide area.

[0138] Specifically, the predicted landslide mask is compared pixel by pixel with the landslide mask generated from the landslide outline label. The number of pixels predicted as landslides and actually landslides, the number of pixels predicted as landslides but not actually landslides, and the number of pixels predicted as not landslides but actually landslides are counted. The number of incorrectly predicted pixels (i.e., the sum of the number of pixels predicted as landslides but not actually landslides and the number of pixels predicted as not landslides but actually landslides) is divided by the total number of pixels included in the statistics. The resulting ratio is the mask deviation metric. The smaller this value, the higher the degree of overlap between the predicted mask and the real mask, and the more accurate the delineation.

[0139] S224, the target box deviation sub-metric, the category deviation sub-metric, and the mask deviation sub-metric are weighted and calculated to obtain the first deviation metric.

[0140] Specifically, a first weighting coefficient can be assigned to the target bounding box deviation sub-metric, a second weighting coefficient to the category deviation sub-metric, and a third weighting coefficient to the mask deviation sub-metric. All three weighting coefficients are positive values. The larger the weighting coefficient, the higher the importance of the sub-metric in the total deviation. Multiply the target bounding box deviation sub-metric by the first weighting coefficient, the category deviation sub-metric by the second weighting coefficient, and the mask deviation sub-metric by the third weighting coefficient, and add the three products together. The resulting value is the first deviation metric. The first deviation metric can comprehensively reflect the overall performance of the patch extraction unit in terms of localization, category judgment, and region delineation. The smaller the value, the higher the overall consistency between the prediction results of the patch extraction unit and the actual annotations.

[0141] S23, perform edge extraction on the predicted landslide mask to obtain a predicted boundary map, determine a second deviation metric based on the difference between the predicted boundary map and the boundary label map, and use the second deviation metric as the boundary deviation metric.

[0142] Among them, the predicted boundary map refers to the image extracted from the predicted landslide mask output by the patch extraction unit, which only represents the boundary location of the predicted landslide area. The pixels at the boundary location in the image are highlighted, and the remaining pixels are the background. The second deviation metric is a numerical index calculated by comparing the predicted boundary map with the generated boundary label map, which reflects the degree of difference between the two at the boundary location. The second deviation metric is the boundary deviation metric, which is used to measure the accuracy of the patch extraction unit in judging the landslide boundary location.

[0143] Understandably, a predicted boundary map can be extracted from the predicted landslide mask and compared with a pre-built boundary label map. The boundary label map is a reference image that identifies the location of the actual landslide boundary. By comparing the predicted boundary map and the boundary label map pixel by pixel, the deviation between the predicted boundary and the actual boundary can be calculated. This deviation is then introduced into the parameter configuration process as an independent boundary deviation metric. This allows the patch extraction unit to reduce not only the overall regional differences but also the differences in boundary positions when adjusting parameters, thereby improving both the regional recognition capability and the accuracy of boundary delineation.

[0144] Specifically, edge extraction is performed on the obtained predicted landslide mask to identify the pixels at the boundary between the landslide area and the non-landslide area in the mask, generating a predicted boundary map. In the predicted boundary map, only the pixels at the boundary position are highlighted, and the remaining pixels are background. The predicted boundary map is compared with the generated boundary label map, and the difference between the two is calculated. The smaller the difference, the better the predicted boundary matches the real boundary. The calculated difference value is used as the second deviation metric, which is the boundary deviation metric used in subsequent parameter adjustments.

[0145] See Figure 3 This is a schematic diagram of the construction process of boundary-aware constraints provided by an embodiment of the present invention. The complete construction logic of the above-mentioned boundary label generation and boundary deviation measurement is as follows: Figure 3 As shown, the whole is divided into two parallel branches, which eventually converge to form the boundary constraints. The upper branch refers to the real boundary generation link, that is, taking the real landslide mask M as input, and after processing by the boundary extraction operator, the real boundary map B is obtained. gt As a benchmark for boundary constraints, the real boundary map accurately marks the standard location of the landslide edge and serves as the basis for judging the accuracy of the predicted boundary. The lower branch refers to the predicted boundary generation link, that is, the predicted landslide mask M output by the patch extraction unit. pred As input, after processing by the boundary extraction operator with the same rules as the upper branch, the predicted boundary map B is obtained. pred This reflects the current identification results of the landslide boundary by the patch extraction unit. The actual boundary map and the predicted boundary map output by the two branches are synchronously input into the boundary deviation calculation module. The difference is calculated through two dimensions: one is pixel-level comparison, that is, comparing the numerical difference between the predicted boundary and the actual boundary pixel by pixel to obtain the boundary pixel deviation; the other is regional overlap calculation, that is, calculating the overlap ratio between the predicted boundary region and the actual boundary region to obtain the boundary overlap deviation. The two types of deviations are weighted and combined according to preset weights to obtain the final boundary deviation metric. This boundary deviation metric will participate in the parameter adjustment of the patch extraction unit together with the regional deviation metric.

[0146] In some embodiments, step S23, "determining a second deviation metric based on the difference between the predicted boundary map and the boundary label map," includes the following steps:

[0147] S231, compare the predicted values ​​of each pixel in the predicted boundary map with the actual values ​​of the corresponding pixels in the boundary label map to obtain a pixel difference measure between the predicted boundary map and the boundary label map.

[0148] Among them, the pixel prediction value refers to the value at each pixel position in the prediction boundary map, the pixel true value refers to the value at each pixel position in the boundary label map, and the pixel difference metric is a numerical value that reflects the overall difference between the two images after comparing the predicted values ​​of all pixel positions in the prediction boundary map with the true values ​​of the corresponding pixel positions in the boundary label map pixel by pixel.

[0149] It is understandable that both the predicted boundary map and the boundary label map are images. The most direct way to determine whether the two are consistent is to compare them pixel by pixel. At each pixel position, the smaller the difference between the predicted value and the true value, the more accurate the predicted boundary is at that position. By summarizing the differences at all pixel positions, we can obtain the statistical value of the overall difference between the two images. This pixel-by-pixel comparison method can accurately detect the deviation between the predicted boundary and the true boundary at each local position.

[0150] Specifically, each pixel position in the predicted boundary map is traversed, and the predicted pixel value at that position is obtained. At the same time, the actual pixel value at the same position in the boundary label map is obtained, and the difference between the two is calculated. The larger the difference, the greater the deviation between the predicted boundary and the actual boundary at that position. After traversing all pixel positions, the differences of all pixel positions are summarized and statistically analyzed. When summarizing, the sum or average of the differences of all pixel positions is taken. The summary result is the pixel difference measure between the predicted boundary map and the boundary label map.

[0151] S232, calculate the overlap ratio between the predicted boundary map and the boundary label map to obtain the boundary overlap metric.

[0152] The overlap ratio refers to the ratio between the number of overlapping pixels in the predicted boundary map and the pixel region marked as the boundary in the boundary label map, and the total number of pixels in both maps. The boundary overlap metric is a value calculated based on the overlap ratio to measure the overall degree of overlap between the predicted boundary and the true boundary. A higher value indicates a better degree of overlap.

[0153] Understandably, pixel difference metric measures local differences, focusing on the specific gap between the predicted and actual values ​​at each pixel location. Overlap ratio, on the other hand, supplements this from the perspective of overall overlap, focusing on how much the predicted boundary region overlaps with the actual boundary region in space. The higher the overlap ratio of the two boundary maps, the more consistent the predicted boundary and the actual boundary are in overall spatial distribution. Using pixel difference metric or overlap ratio alone reflects the accuracy of the boundary from one perspective. Combining the two can provide a more comprehensive assessment of the degree of agreement between the predicted boundary and the actual boundary.

[0154] Specifically, the number of pixels marked as boundaries in the predicted boundary map, the number of pixels marked as boundaries in the boundary label map, and the number of overlapping pixels marked as boundaries in both maps are counted. The overlap ratio is obtained by dividing the number of overlapping pixels by the sum of the number of boundary pixels in the predicted boundary map and the number of boundary pixels in the boundary label map. This overlap ratio can be used as a boundary overlap metric. The higher the value of the boundary overlap metric, the better the overlap between the predicted boundary and the true boundary.

[0155] S233, Based on the pixel difference metric and the boundary coincidence metric, determine the second deviation metric.

[0156] Understandably, pixel difference measure reflects boundary accuracy from the perspective of local pixel-by-pixel difference, focusing on the fine deviation of each pixel position, while boundary coincidence measure reflects boundary accuracy from the perspective of overall overlap ratio, focusing on the degree of spatial overlap of boundary regions. Combining the two to determine the second deviation measure can more comprehensively reflect the degree of deviation between the predicted boundary and the true boundary, making the boundary optimization in the parameter configuration process more thorough.

[0157] Specifically, the obtained pixel difference measure and the obtained boundary coincidence measure are combined for calculation. The combined calculation method can be to assign corresponding coefficients to the pixel difference measure and the boundary coincidence measure respectively and then add them together, or to substitute the two into a preset combined calculation formula for processing. The combined calculation result is used as the second deviation measure. The smaller the second deviation measure, the closer the predicted boundary is to the true boundary.

[0158] S24, based on the first deviation metric and the boundary deviation metric, the internal parameters of the patch extraction unit are iteratively adjusted until the preset conditions are met, and the configured patch extraction unit is obtained.

[0159] Among them, the internal parameters refer to a set of adjustable values ​​in the patch extraction unit that determine how it processes the input image. Iterative adjustment refers to the cyclical process of repeatedly executing input samples, calculating deviations, adjusting parameters, and re-verifying. Each round of adjustment is based on the deviation measurement results of the previous round to fine-tune the parameters, gradually approaching the optimal parameter values. The preset conditions refer to the pre-set criteria for determining whether the parameter adjustment can be stopped, which is usually the change in deviation measurement being less than a certain threshold. The configured patch extraction unit refers to the patch extraction unit whose internal parameters have become stable and meet the preset conditions after multiple rounds of iterative adjustment. This unit has the ability to accurately identify landslide areas and accurately delineate landslide boundaries from remote sensing images.

[0160] Understandably, the internal parameters of a patch extraction unit cannot reach an ideal state through a single adjustment. Multiple rounds are needed to gradually approach the optimal value. In each round of adjustment, the direction and magnitude of parameter adjustment need to be determined based on both the first deviation metric and the boundary deviation metric. If based on the first deviation metric, parameter adjustment will ignore boundary accuracy. If based on the boundary deviation metric, parameter adjustment may deviate from the goal of region identification. By combining the two in a weighted manner, parameter adjustment can simultaneously take into account both region accuracy and boundary accuracy. Iterative adjustment requires a clear stopping condition; otherwise, it will continue indefinitely. When the magnitude of the deviation metric change is very small, it indicates that the parameters have basically stabilized, and further adjustment will not yield significant benefits. At this point, iteration can be stopped, and the current patch extraction unit can be used as the configured unit for subsequent landslide patch extraction tasks.

[0161] Specifically, the first deviation metric and the boundary deviation metric can be used together as the basis for parameter adjustment. One round of adjustment is performed on the internal parameters of the patch extraction unit. After adjustment, the adjustment effect is re-examined using sample images. The results of the examination are used to determine whether the preset conditions are met. If they are met, the adjustment is stopped and the current patch extraction unit is used as the configured patch extraction unit. If they are not met, the next round of adjustment is performed. This process is repeated until the preset conditions are met.

[0162] Based on the above embodiments, step S24 can be implemented in the following ways:

[0163] S241, the first deviation measure and the boundary deviation measure are weighted and combined to obtain a comprehensive deviation measure. Based on the comprehensive deviation measure, the internal parameters of the patch extraction unit are adjusted to obtain the adjusted patch extraction unit.

[0164] Among them, the comprehensive deviation metric refers to the comprehensive value obtained by weighting the first deviation metric and the boundary deviation metric according to their respective weight coefficients. This value can simultaneously reflect the overall performance of the patch extraction unit in both region recognition and boundary delineation.

[0165] Understandably, the first deviation metric and the boundary deviation metric measure the performance of the patch extraction unit in two different dimensions. By weighting and combining the two into a comprehensive deviation metric, the optimization objective can be unified. Each adjustment can be based on a comprehensive index to determine the direction and magnitude of the adjustment. The weight coefficient can be adjusted according to the emphasis on regional accuracy and boundary accuracy in actual applications. When higher boundary accuracy is required, a larger weight can be assigned to the boundary deviation metric.

[0166] Specifically, a corresponding weight coefficient is assigned to the first deviation metric and a corresponding weight coefficient is assigned to the boundary deviation metric. The first deviation metric is multiplied by its corresponding weight coefficient, and the boundary deviation metric is multiplied by its corresponding weight coefficient. The two products are then added together, and the resulting value is the comprehensive deviation metric. Based on this comprehensive deviation metric, the internal parameters of the patch extraction unit are adjusted. The goal of the adjustment is to gradually reduce the comprehensive deviation metric. After the adjustment, a patch extraction unit with updated internal parameters is obtained.

[0167] S242, the sample images in the sample image set are input again into the adjusted patch extraction unit to calculate a new comprehensive deviation metric.

[0168] Understandably, after each round of parameter adjustment, the internal parameters of the patch extraction unit change, and its landslide identification and boundary delineation results will also change accordingly. Only by reprocessing the sample images with the adjusted unit can we verify whether the adjustment has truly achieved the desired improvement effect. Without re-verification, it is impossible to know whether the parameter adjustment is in the right or wrong direction.

[0169] Specifically, the sample image is input again into the adjusted patch extraction unit to obtain the adjusted prediction result. Based on the adjusted prediction result, the first deviation metric and the boundary deviation metric are recalculated, and the comprehensive deviation metric is recalculated according to the weighted combination method in step S241 to obtain the adjusted comprehensive deviation metric, i.e., the new comprehensive deviation metric.

[0170] S243, when the change between the new comprehensive deviation metric and the comprehensive deviation metric before one round of adjustment is less than a preset threshold, stop the iterative adjustment and use the current patch extraction unit as the configured patch extraction unit.

[0171] Among them, the change range refers to the absolute value of the difference between the new comprehensive deviation metric and the comprehensive deviation metric before adjustment, and the preset threshold refers to the pre-set critical value used to determine whether parameter adjustment can be stopped.

[0172] Understandably, in the early stages of iterative adjustments, parameter adjustments usually bring about significant improvements, with a large decrease in the overall deviation metric. As the number of adjustment rounds increases, the parameters gradually approach the optimal value, and the improvement brought by each round of adjustment becomes smaller and smaller. When the improvement is already very small, the benefit of continuing to adjust is very low, and it may cause the parameters to fluctuate around the optimal value without stabilizing. Therefore, a judgment criterion for stopping adjustment can be set, namely a preset threshold.

[0173] Specifically, the absolute value of the difference between the calculated new comprehensive deviation metric and the comprehensive deviation metric before this round of adjustment is used as the change range. The change range is compared with a preset threshold. If the change range is less than the preset threshold, it means that the parameter has stabilized. At this time, the iterative adjustment can be stopped, and the current patch extraction unit can be used as the configured patch extraction unit.

[0174] S244, if the change magnitude is greater than or equal to a preset threshold, continue to perform the next round of parameter adjustment based on the new comprehensive deviation metric until the change magnitude meets the preset threshold.

[0175] Understandably, when the change is still greater than the preset threshold, it means that the improvement brought about by this round of adjustment is quite significant, and there is still room for further optimization of the parameters. At this time, the adjustment should not be stopped, but the state after this round of adjustment should be taken as the starting point to continue to execute the next round of adjustment.

[0176] Specifically, when the calculated change is greater than or equal to the preset threshold, starting from the new comprehensive deviation metric obtained in step S242, return to step S241 to continue adjusting the internal parameters of the patch extraction unit in the next round. This process is repeated until the change after a certain round of adjustment is less than the preset threshold, thus satisfying the stopping condition.

[0177] S3, acquire the remote sensing image to be processed, divide the remote sensing image to be processed into sub-image blocks based on different image block specifications, input the sub-image blocks into the configured patch extraction unit, and obtain the sub-probability map corresponding to each image block specification.

[0178] Among them, the remote sensing image to be processed refers to the complete remote sensing image that needs to be extracted for landslide patches; the image block specification refers to the size specification of the sub-image blocks used to divide the image; the sub-image block refers to the local image region divided from the remote sensing image to be processed according to a certain image block specification; and the sub-probability map refers to the landslide probability distribution map corresponding to the block specification formed by stitching together the processing results of all sub-image blocks belonging to the same image block specification according to their original positions after processing by the patch extraction unit.

[0179] Understandably, remote sensing images typically cover large areas, with the overall image size far exceeding the image size that the patch extraction unit can directly process. Therefore, the entire image cannot be directly input into the patch extraction unit; it needs to be divided into several sub-image blocks for separate processing. Different image block sizes are used because the scale of landslides in remote sensing images can vary significantly. Small landslides may only occupy tens of pixels, while large landslides may span thousands of pixels. If a single, small block size is used, the geographical area covered by each sub-image block is limited. Large landslides would be divided into multiple sub-image blocks, and each sub-image block can only present a localized portion of the landslide. In some areas, the image patch extraction unit cannot see the entire landslide, resulting in fragmented and incomplete outlines of large landslide patches after stitching. If a single and large-sized block specification is used, each sub-image block covers a large geographical area, and the proportion of pixels occupied by small landslides in the sub-image blocks is too small. The image patch extraction unit has difficulty effectively capturing the features of these small targets, which can easily lead to missed detections. Therefore, at least two different image block specifications can be preset to divide and process the same remote sensing image to be processed. The smaller specification has a stronger ability to distinguish the details of small landslides, while the larger specification can completely include large landslides within the field of view of a single sub-image block.

[0180] Specifically, the complete remote sensing image for which landslide patch extraction is required is obtained as the remote sensing image to be processed. Then, at least two different image block specifications can be preset. For each image block specification, the remote sensing image to be processed is divided into several sub-image blocks according to the size of the specification. The sub-image blocks corresponding to each specification are input one by one to the patch extraction unit with completed parameter configuration. The unit processes each sub-image block to obtain the sub-probability map corresponding to each image block specification.

[0181] Based on the above embodiments, step S3 can be implemented in the following ways:

[0182] S31, at least two different image block specifications are preset. For each image block specification, the remote sensing image to be processed is divided based on the corresponding image block division method. There is an overlapping area between adjacent sub-image blocks, and a set of sub-image blocks corresponding to the image block specification is obtained.

[0183] Among them, image block specification refers to the size of the sub-image blocks used to divide the image. Different image block specifications correspond to different sub-image block sizes. Image division method refers to the method of determining the specific location and range of each sub-image block on the remote sensing image to be processed according to the size specified by the image block specification. Overlapping area refers to the part of the area shared by two adjacent sub-image blocks. This area belongs to two adjacent sub-image blocks at the same time and can be used as a buffer zone for stitching and fusion in subsequent processing. Sub-image block set refers to the set of all sub-image blocks divided under a certain image block specification after the remote sensing image to be processed is divided according to a certain image block specification.

[0184] Understandably, pre-setting at least two different image block sizes is to allow each size to take advantage of its strengths. Smaller sizes cover a smaller geographical area per sub-image block, preserving sufficient image details, which is beneficial for the identification of small landslides. Larger sizes cover a larger geographical area per sub-image block, which can completely contain large landslides in a single sub-image block, avoiding fragmentation. During the division, overlapping areas between adjacent sub-image blocks are intended to avoid the landslide target being abruptly cut in half at the boundary of the sub-image blocks, resulting in obvious seam marks during subsequent stitching. The overlapping areas can provide a certain buffer for subsequent stitching, allowing the prediction results of adjacent sub-image blocks to transition smoothly in the overlapping part.

[0185] Specifically, based on the size of the remote sensing image to be processed and the scale distribution of the landslide target, at least two sub-image block sizes can be predetermined as different image block specifications. For each image block specification, starting from the upper left corner of the remote sensing image to be processed, the sub-image block size can be used as the basic unit and the image can be moved to the right and down gradually with a preset step size, according to the sub-image block size specified in the specification. The position and range of each sub-image block in the image can be determined in turn. During the division, some areas of adjacent sub-image blocks overlap with each other. That is, the current sub-image block shares some pixel columns with its right-side adjacent sub-image block in the horizontal direction and shares some pixel rows with its lower-side adjacent sub-image block in the vertical direction. After the division is completed, all sub-image blocks under this specification form the sub-image block set corresponding to this specification.

[0186] S32, input the set of sub-image blocks corresponding to each of the image block specifications into the configured patch extraction unit to obtain the landslide prediction results of each of the sub-image blocks.

[0187] The landslide prediction result refers to an image of the same size as the sub-image patch, which is output by the configured patch extraction unit after a single sub-image patch is input into it. The value of each pixel in the image represents the probability that the location belongs to a landslide, and the higher the value, the more likely the location is to belong to a landslide area.

[0188] Understandably, the previous steps have divided the remote sensing image to be processed into multiple sets of sub-image blocks according to different specifications. Each sub-image block in the set is an independent local image region. The patch extraction unit has also completed parameter configuration and has the ability to identify landslide areas from the image and give a probability judgment. However, the image size it can directly process is fixed. Therefore, each sub-image block needs to be input into the unit one by one for processing. After each sub-image block is input, the patch extraction unit will analyze the position of each pixel in the sub-image block according to its internal parameters and output a probability value to represent its judgment on the possibility that the position belongs to a landslide. Different image block specifications correspond to different sub-image block sizes. Larger sub-image blocks cover a larger area of ​​the ground, while smaller sub-image blocks cover a smaller area of ​​the ground but have clearer details.

[0189] Specifically, for each set of sub-image blocks corresponding to each image block specification, each sub-image block in the set is input one by one into the patch extraction unit with completed parameter configuration. The patch extraction unit can perform pixel-by-pixel analysis processing on each input sub-image block and output an image with the same size as the sub-image block as the landslide prediction result. The value of each pixel in the landslide prediction result is the probability that the unit judges that the pixel belongs to a landslide. After all sub-image blocks of the same image block specification have been processed, the landslide prediction result corresponding to each sub-image block under that specification is obtained.

[0190] S33, the landslide prediction results of each sub-image block belonging to the same image block specification are stitched together according to the original position of each sub-image block in the remote sensing image to be processed to obtain the sub-probability map corresponding to each image block specification.

[0191] The original location refers to the position range of the sub-image patch in the remote sensing image to be processed before it is divided. It is usually represented by the row and column number of the top left pixel of the sub-image patch in the whole image.

[0192] Understandably, the landslide prediction result obtained from each sub-image patch is a probability map of the same size as the sub-image patch, but it only covers a part of the remote sensing image to be processed. In order to obtain a complete probability map covering the entire image, these scattered local probability maps can be reassembled according to their original positions in the image, just like putting the pieces of a jigsaw puzzle back in their original positions to restore a complete picture. During the stitching, the original position information of each sub-image patch is the only basis for determining where each local probability map should be placed.

[0193] Specifically, for the same image block specification, the landslide prediction results corresponding to each sub-image block under that specification are obtained. Each landslide prediction result is a probability map. The original position information of each sub-image block in the remote sensing image to be processed is obtained. According to the original position of each sub-image block, its corresponding probability map is placed in the corresponding position in a blank image of the same size as the remote sensing image to be processed. After all the probability maps of the sub-image blocks are placed, a complete probability map covering the entire range of the remote sensing image to be processed is formed. This probability map is the sub-probability map corresponding to the image block specification. The above stitching operation is performed on each image block specification to obtain the sub-probability map corresponding to each image block specification.

[0194] S4. The sub-probability maps are weighted and fused to generate a comprehensive landslide probability map. The comprehensive landslide probability map is then transformed to obtain landslide patch data, which is then output.

[0195] Among them, the comprehensive landslide probability map refers to the comprehensive landslide probability distribution map obtained by weighted fusion of sub-probability maps corresponding to different image block sizes. The value of each pixel in the map comprehensively reflects the probability that the location belongs to a landslide. The conversion process refers to a series of processing operations that transform the landslide probability map from a pixel probability distribution form into patch data with clear boundaries and geographic coordinates. The landslide patch data refers to the landslide patch result data with geospatial reference information obtained after conversion processing.

[0196] Understandably, sub-probability maps obtained from different image block sizes may give different judgments on the probability that the same ground location belongs to a landslide. Smaller sizes have a stronger ability to distinguish details of small landslides and usually give more reliable probability values ​​for small landslide areas. Larger sizes have more complete outline coverage of large landslides and usually give more reliable probability values ​​for large landslide areas. By weighted fusion, the probability judgments of each size can be combined to obtain a landslide probability map that is more reliable than any single size.

[0197] Specifically, the sub-probability maps corresponding to each image block size can be weighted and fused to generate a comprehensive landslide probability map. The comprehensive landslide probability map is then transformed into patch data with clear boundaries by converting the probability distribution image into patch data with clear boundaries. Combined with the geospatial reference information carried by the remote sensing image itself, landslide patch data with geospatial reference is obtained and output.

[0198] Based on the above embodiments, step S4 can be implemented in the following ways:

[0199] S41, assign corresponding fusion weights to each of the image block specifications, and perform a weighted average of the pixel values ​​of the same geographical location in each of the sub-probability maps according to the fusion weights to generate a comprehensive landslide probability map.

[0200] Among them, the fusion weight refers to a numerical coefficient assigned to each image block specification, which represents the reliability of the probability value in the sub-probability map corresponding to that specification. The larger the weight, the more reliable the judgment of that specification is. Same geographical location refers to the pixel position of the same ground location on the remote sensing image to be processed in different sub-probability maps. The comprehensive landslide probability map is a new probability map obtained by weighting and fusing the sub-probability maps of each specification according to their respective weights.

[0201] Understandably, by assigning fusion weights to each specification and performing weighted fusion, the judgments of each specification on the location can be combined. Specifications with higher credibility will have a larger weight in the combined result, thus obtaining a more reliable landslide probability value than any single specification.

[0202] Specifically, based on the reliability of each image block specification in landslide recognition, a fusion weight value is determined for each specification. According to the fusion weight corresponding to each specification, the sub-probability maps are weighted and fused. The probability values ​​of each specification are combined according to their weight ratio. After fusion, a new probability map is obtained, which is the comprehensive landslide probability map.

[0203] S42, perform threshold segmentation on the comprehensive landslide probability map to obtain a binary landslide mask.

[0204] Threshold segmentation refers to comparing the probability value of each pixel in the comprehensive landslide probability map with a preset segmentation threshold, and re-determining the pixel as a landslide pixel or a non-landslide pixel based on the comparison result. Binarized landslide mask refers to the image obtained after threshold segmentation that contains only two values: landslide pixel and non-landslide pixel. Landslide pixel indicates that the location is determined to be a landslide area, and non-landslide pixel indicates that the location is determined to be a non-landslide area.

[0205] Understandably, in a comprehensive landslide probability map, each pixel's value is a probability value, representing the likelihood of that location being a landslide. While this probability map contains relatively rich information, it cannot directly extract specific landslide boundary patterns. To determine whether a location is a landslide, a clear segmentation threshold needs to be set as the judgment standard. That is, pixels with a probability value higher than or equal to the set segmentation threshold can be identified as landslide pixels, while pixels with a probability value lower than the set segmentation threshold can be identified as non-landslide pixels. After threshold segmentation, the probability map is transformed into a binary image with only two possible values, and the landslide area and the non-landslide area have a clear boundary.

[0206] Specifically, a segmentation threshold can be set, and the probability value of each pixel in the comprehensive landslide probability map can be compared with the segmentation threshold one by one. Pixels with a probability value greater than or equal to the segmentation threshold are identified as landslide pixels, and pixels with a probability value less than the segmentation threshold are identified as non-landslide pixels. After all pixels have been processed, a binary image composed of the two values ​​of landslide pixels and non-landslide pixels is obtained. This image is the binarized landslide mask.

[0207] S43, sequentially perform closing operation, hole filling operation, opening operation and connected region area filtering operation on the binarized landslide mask to obtain the refined landslide mask.

[0208] Among them, the closing operation refers to the process of first performing a dilation operation on the landslide area in the binarized landslide mask, and then performing an erosion operation, which is used to bridge the small fractures and gaps on the boundary of the landslide patch. The hole filling operation refers to the process of filling the closed non-landslide area surrounded by landslide pixels in the binarized landslide mask with landslide pixels, which is used to fill the holes inside the patch. The opening operation refers to the process of first performing an erosion operation on the landslide area in the binarized landslide mask, and then performing a dilation operation, which is used to remove burrs and isolated small noise points at the edge of the patch. The connected component area filtering operation refers to the process of calculating the number of pixels contained in each independent connected region composed of landslide pixels in the binarized landslide mask, and removing the connected regions whose number of pixels is less than a preset minimum area threshold, which is used to remove scattered small false detection patches. The refined landslide mask refers to the landslide mask image obtained after the above series of processing, which has continuous and complete boundaries, no holes inside, and no small noise spots.

[0209] It is understandable that the binary landslide mask obtained by threshold segmentation may have various morphological defects. For example, there may be small cracks at the boundary of the landslide area, causing the originally connected areas to be broken. There may be cavities inside the landslide area, burr-like protrusions at the edge of the landslide area, and some isolated noise points with small areas caused by noise in the image. If these defects are not processed, they may affect the quality of subsequent contour extraction, resulting in incomplete boundaries, gaps inside, or a large number of fragmented landslide patches in the final output. Closing operations can bridge the small fractures on the boundary, hole filling can fill the internal cavities, opening operations can remove edge burrs and isolated noise points, and connected component area filtering can remove noise patches with too small an area. The above operations are performed in the order of closing operations and hole filling first, followed by opening operations and area filtering, which can gradually improve the mask quality and finally obtain a refined landslide mask with complete morphology and clear boundaries.

[0210] Specifically, for the obtained binarized landslide mask, a closing operation is first performed, and an expansion operation is performed on the landslide area in the mask to expand the boundary outward and thus close the small cracks. Then, an erosion operation is performed to shrink the expanded landslide area inward back to its original size. A hole-filling operation is performed on the result after the closing operation to identify the closed non-landslide areas surrounded by landslide pixels in the mask. These non-landslide pixels in the closed areas are re-identified as landslide pixels. An opening operation is performed on the result after the hole-filling operation. First, an erosion operation is performed on the landslide area to remove the small burrs and isolated noise at the edges. Then, an expansion operation is performed to restore the main shape of the landslide area. A connected component area filtering operation is performed on the result after the opening operation. Each independent connected region composed of landslide pixels in the mask is traversed, and the number of pixels contained in each connected region is calculated. Connected regions with a number of pixels less than a preset minimum area threshold are completely removed. After the above steps are performed in sequence, a refined landslide mask can be obtained.

[0211] S44, extract the outline of each landslide patch in the refined landslide mask, combine it with the georeferenced information of the remote sensing image to be processed, convert the outline pixel coordinates into geographic coordinates, obtain the landslide patch data and output it.

[0212] Among them, georeferenced information refers to the parameter information recorded in the remote sensing image to be processed, which is used to establish the correspondence between image pixel coordinates and real-world geographic coordinates; contour pixel coordinates refer to the position of each vertex on the extracted landslide patch contour in the image, represented by pixel row number and column number; and geographic coordinates refer to the position coordinates in the real-world coordinate system obtained after converting the contour pixel coordinates through georeferenced information.

[0213] Understandably, in the refined landslide mask obtained, the boundary of each landslide patch is still represented by pixel coordinates. Pixel coordinates are only meaningful within a specific image; without the image, the location of the patch in the real world cannot be determined. In order to use the landslide patch data for practical geographic information analysis, disaster assessment, and emergency decision-making, the boundary of each patch can be converted from pixel coordinates to real-world geographic coordinates. Since the remote sensing image to be processed has already recorded geographic reference information such as the image projection method and affine transformation parameters during acquisition, this information provides the basis for the conversion from pixel coordinates to geographic coordinates. After coordinate conversion, each landslide patch can have real-world location information and can be overlaid and analyzed with other geographic data in GIS software, becoming standardized result data that can be used directly.

[0214] Specifically, the refined landslide mask can be used to identify each independent landslide patch region composed of landslide pixels, track and record the pixel positions passed by the boundary along the outer boundary of each patch, form the outline pixel coordinate sequence of the patch in sequence, obtain the georeferenced information recorded in the remote sensing image to be processed, use this information to convert each pixel coordinate on the outline of each landslide patch into the corresponding georeferenced coordinates, after the conversion is completed, organize the georeferenced coordinate sequence of each patch according to the preset data format, obtain the landslide patch data and output it.

[0215] See Figure 4 This is a schematic diagram of a boundary-aware remote sensing image landslide patch extraction device provided in an embodiment of the present invention. The data processing system of this boundary-aware remote sensing image landslide patch extraction device includes:

[0216] The construction module is used to construct a sample image set based on the acquired first sample image and second sample image, and to perform morphological processing on the landslide contour label corresponding to the first sample image to generate a boundary label map.

[0217] The configuration module is used to configure the parameters of the patch extraction unit based on the sample image set and the boundary deviation metric corresponding to the boundary label map, so as to obtain the configured patch extraction unit.

[0218] The segmentation module is used to acquire the remote sensing image to be processed, divide the remote sensing image to be processed into sub-image blocks based on different image block specifications, and input the sub-image blocks into the configured patch extraction unit to obtain sub-probability maps corresponding to each image block specification.

[0219] The conversion module is used to perform weighted fusion of the sub-probability maps to generate a comprehensive landslide probability map, and to perform conversion processing on the comprehensive landslide probability map to obtain landslide patch data and output it.

[0220] Figure 4 The apparatus of the illustrated embodiment can be used to perform corresponding actions. Figure 1 The steps in the method embodiments shown are implemented in a similar manner and have similar technical effects, and will not be repeated here.

[0221] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for extracting landslide patches from remote sensing images based on boundary awareness, characterized in that, include: A sample image set is constructed based on the first and second sample images obtained. The landslide contour labels corresponding to the first sample image are morphologically processed to generate a boundary label map. Based on the sample image set and the boundary deviation metric corresponding to the boundary label map, the parameters of the patch extraction unit are configured to obtain the configured patch extraction unit, including: The sample images in the sample image set are input into the patch extraction unit to obtain the predicted landslide mask output by the patch extraction unit; A first deviation metric is determined based on the difference between the predicted landslide mask and the landslide mask corresponding to the landslide outline label; Edge extraction is performed on the predicted landslide mask to obtain a predicted boundary map. Based on the difference between the predicted boundary map and the boundary label map, a second deviation metric is determined and used as the boundary deviation metric. Based on the first deviation metric and the boundary deviation metric, the internal parameters of the patch extraction unit are iteratively adjusted until the preset conditions are met, thus obtaining the configured patch extraction unit. The remote sensing image to be processed is acquired, and the image is divided into sub-image blocks based on different image block specifications. The sub-image blocks are then input into the configured patch extraction unit to obtain sub-probability maps corresponding to each image block specification. The sub-probability maps are weighted and fused to generate a comprehensive landslide probability map. This comprehensive landslide probability map is then transformed to obtain landslide patch data, which is then output. Assign corresponding fusion weights to each of the image block specifications, and perform a weighted average of the pixel values ​​of the same geographical location in each of the sub-probability maps according to the fusion weights to generate a comprehensive landslide probability map; Threshold segmentation is performed on the comprehensive landslide probability map to obtain a binarized landslide mask; The binarized landslide mask is sequentially subjected to closing operation, hole filling operation, opening operation, and connected component area filtering operation to obtain a refined landslide mask; The contours of each landslide patch in the refined landslide mask are extracted, and combined with the georeferenced information of the remote sensing image to be processed, the contour pixel coordinates are converted into geographic coordinates to obtain landslide patch data and output it.

2. The method according to claim 1, characterized in that, The construction of the sample image set based on the acquired first and second sample images includes: The acquired remote sensing image slice containing landslide targets and labeled with landslide annotation information is used as the first sample image. The landslide annotation information is a landslide outline label or a landslide mask label. The remote sensing image slices that do not contain the landslide target but contain land cover types that are easily confused with the landslide target are used as the second sample images; A sample image set is constructed based on the first sample image and the second sample image; When constructing the sample image set, the ratio of the number of first sample images to the number of second sample images is controlled to be within a preset ratio range; When the landslide annotation information is a landslide mask label, before performing morphological processing on the landslide contour label corresponding to the first sample image, the method further includes: Connectivity analysis was performed on the landslide mask labels to extract the contour vertex coordinates of each landslide patch; The coordinates of the contour vertices are normalized to obtain the landslide contour label.

3. The method according to claim 1, characterized in that, The step of performing morphological processing on the landslide contour label corresponding to the first sample image to generate a boundary label map includes: The polygonal region enclosed by the coordinates of each vertex in the landslide outline label is filled with pixels to generate a landslide mask; A gradient convolution kernel is applied to the landslide mask to calculate the gradient magnitude at each pixel location, and a boundary label map is generated after normalization. Alternatively, an expansion operation can be performed on the landslide mask to obtain an expanded mask, and an corrosion operation can be performed on the landslide mask to obtain a corrosion mask; The boundary label map is obtained by subtracting the expansion mask from the corrosion mask.

4. The method according to claim 1, characterized in that, The step of determining a first deviation metric based on the difference between the predicted landslide mask and the landslide mask corresponding to the landslide contour label includes: Obtain the predicted target box output by the patch extraction unit, and determine the target box deviation sub-metric based on the positional difference between the predicted target box and the real target box corresponding to the landslide contour label; Obtain the predicted category confidence score output by the patch extraction unit, and determine the category bias sub-measure based on the difference between the predicted category confidence score and the true category label; Obtain the regional overlap difference between the predicted landslide mask and the landslide mask corresponding to the landslide outline label, and determine the mask deviation sub-metric; The target box deviation sub-metric, the category deviation sub-metric, and the mask deviation sub-metric are weighted and calculated to obtain the first deviation metric.

5. The method according to claim 1, characterized in that, The step of determining a second deviation metric based on the difference between the predicted boundary map and the boundary label map includes: The predicted values ​​of each pixel in the predicted boundary map are compared with the actual values ​​of the corresponding pixels in the boundary label map to obtain a pixel difference measure between the predicted boundary map and the boundary label map. Calculate the overlap ratio between the predicted boundary map and the boundary label map to obtain the boundary overlap metric; A second deviation metric is determined based on the pixel difference metric and the boundary coincidence metric.

6. The method according to claim 1, characterized in that, The step of iteratively adjusting the internal parameters of the patch extraction unit based on the first deviation metric and the boundary deviation metric until a preset condition is met, to obtain the configured patch extraction unit, includes: The first deviation measure and the boundary deviation measure are weighted and combined to obtain a comprehensive deviation measure. Based on the comprehensive deviation measure, the internal parameters of the patch extraction unit are adjusted to obtain the adjusted patch extraction unit. The sample images in the sample image set are input again into the adjusted patch extraction unit to calculate a new comprehensive deviation metric. When the change between the new comprehensive deviation metric and the comprehensive deviation metric before one round of adjustment is less than a preset threshold, the iterative adjustment is stopped, and the current patch extraction unit is used as the configured patch extraction unit. If the change is greater than or equal to a preset threshold, the next round of parameter adjustment will continue based on the new comprehensive deviation metric until the change meets the preset threshold.

7. The method according to claim 1, characterized in that, The process of acquiring the remote sensing image to be processed involves dividing the image into sub-image blocks based on different image block specifications, and inputting these sub-image blocks into the configured patch extraction unit to obtain sub-probability maps corresponding to each image block specification, including: At least two different image block specifications are preset. For each image block specification, the remote sensing image to be processed is divided based on the corresponding image block division method. There are overlapping areas between adjacent sub-image blocks, and a set of sub-image blocks corresponding to the image block specification is obtained. The sub-image block sets corresponding to each of the image block specifications are respectively input into the configured image patch extraction unit to obtain the landslide prediction results of each of the sub-image blocks; The landslide prediction results of each sub-image block belonging to the same image block size are stitched together according to the original position of each sub-image block in the remote sensing image to be processed to obtain the sub-probability map corresponding to each image block size.

8. The landslide patch extraction device based on boundary awareness from remote sensing images according to any one of claims 1-7, characterized in that, include: The construction module is used to construct a sample image set based on the acquired first sample image and second sample image, and to perform morphological processing on the landslide contour label corresponding to the first sample image to generate a boundary label map. The configuration module is used to configure the parameters of the patch extraction unit based on the sample image set and the boundary deviation metric corresponding to the boundary label map, so as to obtain the configured patch extraction unit. The segmentation module is used to acquire the remote sensing image to be processed, divide the remote sensing image to be processed into sub-image blocks based on different image block specifications, and input the sub-image blocks into the configured patch extraction unit to obtain sub-probability maps corresponding to each image block specification. The conversion module is used to perform weighted fusion of the sub-probability maps to generate a comprehensive landslide probability map, and to perform conversion processing on the comprehensive landslide probability map to obtain landslide patch data and output it.

Images